Scaling AI Coding Assistant Adoption Across Engineering Organization

hubspot · Tech · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Engineering-wide AI coding adoption

HubSpot needed to move AI coding assistants from cautious experimentation to near-universal engineering adoption while preserving code quality, proving safety with data, tailoring tools to HubSpot's opinionated stack, and turning AI fluency into an organizational baseline.

Constraints:

• Initial adoption required legal, security, engineering, and executive alignment.
• Skeptics worried about production incidents, review burden, cycle time, and quality.
• Adoption slowed after early users and again after passing 60%.
• HubSpot's opinionated stack required generated code to reflect internal architecture, libraries, and best practices.
• Procurement needed to move faster than traditional negotiated agreements.

Architecture

HubSpot scaled AI coding by combining a large GitHub Copilot pilot, executive sponsorship, enablement channels, engineering velocity measurement, and a dedicated Developer Experience AI team. The team then broadened adoption through peer validation, high-level quantitative proof, multiple coding-assistant POCs, a curated local MCP server with default rules and configuration, AI community advocacy, faster procurement, and AI-fluency expectations in hiring.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["HubSpot engineers"] --> B["AI coding tools"]
  C["Developer Experience AI team"] --> B
  C --> D["Metrics and evaluation"]
  C --> E["AI community"]
  B --> F["Local MCP server"]
  F --> G["Internal MCP tool ecosystem"]
  G --> B
  E --> A
  D --> C

Components

Data flow

The DevEx AI team pilots tools, aligns legal/security/procurement, measures adoption and engineering outcomes, curates rules and MCP configuration for HubSpot's stack, and feeds learning back through enablement channels, videos, community posts, and hiring expectations. Engineers use coding assistants and agents, while metrics such as review burden, cycle time, velocity, adoption, and incident rates guide rollout decisions.

Why this architecture

HubSpot treated adoption as an engineering platform problem rather than a license rollout. A dedicated DevEx AI team created leverage by measuring impact, curating context for the internal stack, speeding procurement, building community, and reducing friction as adoption moved from early adopters to skeptics.

Models used: GitHub Copilot · Cursor · OpenAI models · Anthropic models · coding agents

Scale

• HubSpot reached more than 90% AI coding adoption.
  “Once we hit 90% adoption, we made AI fluency a baseline expectation for engineers”article, stated
• Adoption slowed after passing 60%, requiring a new strategy for skeptics and risk-averse users.
  “Adoption slowed again as it increased beyond 60%.”article, stated
• The initial Copilot pilot ran for over two months.
  “We gave teams over two months to try it..”article, stated
• GitHub Copilot cost $19 per business user per month at the time of the initial pilot.
  “Copilot's cost structure of $19/mo/business user at the time.”article, stated
• HubSpot later exposed 400-plus tools across internal, OpenAI, and Anthropic MCP servers.
  “building infrastructure that led to 400+ tools that our agents can leverage across our internal, OpenAI, and Anthropic MCP servers.”article, stated

Alternatives considered (roads not taken)

License-request guardrails for every user

Rejected because: Conservative usage rules slowed adoption once HubSpot had enough data and experience.

Qualitative feedback alone

Rejected because: HubSpot wanted empirical data to combat skepticism and compare tools on merit.

Generic AI tool defaults

Rejected because: HubSpot needed generated code to follow its specific stack and best practices.

Patterns

context-engineering (context)

HubSpot curates rules, local MCP configuration, and deep architectural context so generated code reflects its internal stack.

tool-use-orchestration (orchestration)

HubSpot exposes hundreds of tools through MCP servers that agents can use across internal, OpenAI, and Anthropic contexts.

eval-driven-development (evaluation)

HubSpot uses engineering velocity metrics, incident data, pilot comparisons, and tool evaluation capabilities to guide rollout.

human-in-the-loop-gate (governance)

Adoption relied on peer validation, forums, videos, and engineering community review rather than a top-down mandate alone.

Tradeoffs

Adoption rollout

Chose large team-based pilot with measurement over small individual-only experiment.

Rationale: Including full teams with varied experience let people learn together and produced enough data to validate impact.

Adoption expectation

Chose AI fluency as a baseline expectation over optional specialist skill.

Rationale: After 90% adoption, HubSpot judged AI fluency necessary for engineers' growth and hiring expectations.

Procurement mode

Chose faster month-to-month experimentation over longer-term negotiated agreements before pilots.

Rationale: AI tools change quickly and HubSpot wanted to try new tools without waiting for traditional procurement.

Gotchas (what bit them)

Demand can outgrow adjacent teams that only partly own AI adoption.

Root cause: GitHub-adjacent teams initially drove adoption, but demand exploded and their backlog grew.

Resolution: HubSpot created a dedicated Developer Experience AI team in October 2024.

Late-adoption skeptics need different proof than early adopters.

Root cause: After 60% adoption, remaining engineers had more skepticism, risk aversion, and awareness of tool limitations.

Resolution: HubSpot used peer videos, broad quantitative proof, better tools, curated MCP setup, and AI fluency expectations.

Generic assistants do not know an opinionated internal stack.

Root cause: HubSpot wanted generated code to reflect its architecture, libraries, and practices.

Resolution: The DevEx AI team provided Cursor rules, local MCP configuration, and deeper architectural context.

Evaluation

HubSpot evaluated adoption and safety with engineering velocity methods, qualitative feedback, code review burden, cycle time, before/after velocity comparisons, production incident rates, broad adoption data, and pilot comparisons.

Metrics: AI coding adoption percentage · Code review burden · Cycle time · Velocity before and after adoption · Production incident rates · Pilot feedback · Tool comparison results

Outcomes

• HubSpot reached near-universal AI coding adoption and later passed 90%.
• The company created a dedicated Developer Experience AI team in October 2024.
• Data showed no correlation between AI adoption and production incidents.
• HubSpot made AI fluency a baseline expectation in job descriptions and hiring.
• The adoption work opened the door to Sidekick, coding agents, rapid UI prototyping, and 400-plus MCP tools.

HubSpot reports adoption above 90%, a two-month-plus pilot, GitHub Copilot's $19/month business-user pricing at the time, adoption slowdown after 60%, and 400-plus tools across internal/OpenAI/Anthropic MCP servers.